A Systematic Review of Enhancing CNN Performance in Automated Fabric Defect Detection Through Sampling Techniques for Imbalanced Datasets with the Developed CNN Model

Abstract

Saima Saleem, David Williams and Satya Prakash

In the textile industry, manual fabric inspection poses significant challenges. Incomplete and faulty inspections can compromise both product cost and quality. With the advancements in deep learning, various machine learning algorithms have emerged as successful tools for image classification and analysis tasks. Nevertheless, there are several persistent issues, including the complexity and time-consuming nature of training methods, the requirement for large datasets, and difficulties in achieving generalization. What's needed is an accurate and swift automatic machine learning algorithm suitable for real-time detection in industrial setups. To tackle these challenges, this research successfully developed a straightforward Convolutional Neural Network (CNN) machine learning algorithm.

The algorithm's performance was evaluated on two different image sizes: 150 x 700, and 245 x 345. It became evident that image size significantly influences the model's performance. Additionally, the dataset's inherent imbalance had an adverse impact on the model's performance due to inadequate training and overfitting. To address the issue of imbalanced dataset and enhance the model's performance, various sampling techniques were experimented with. Among these, the CNN model exhibited its most outstanding performance when paired with a smaller image size of 245x345 and when utilizing the SMOTEENN sampling technique. The results demonstrated remarkable accuracy, precision, recall, and F1 score, with values of 98.00%, 98.00%, 98.00%, and 98.00%, respectively. Moreover, the time required for modelling and prediction was impressively low, at 1.57 seconds and 0.09 seconds, respectively. The research also proposed a method to deploy the algorithm and automate the entire quality inspection process within the textile industry.

PDF